Stage 2 / Prompt Toolkit

STAGE 2 · REVIEW · HANDS-ON PROMPT BANK

Stage 2 Prompt Toolkit

Use prompts alongside the seven review steps. Replace bracketed fields, check every suggestion against your protocol and source evidence, then record only verified outputs in the Stage 2 worksheets.

Open Stage 2 workspace →Stage 2 overview Gemini Notebook guide Build a Stage 2 prompt

Human reviewers make final decisions. Do not upload confidential full texts unless permitted by institutional policy. AI suggestions must not be counted as independent reviewer votes. Keep screening, appraisal and extraction decisions in the review system and retain an audit trail.

How to use the prompts

Open the relevant step and tool guide (for example Zotero, Rayyan, Covidence, Gemini or ChatGPT).
Replace every bracketed field; share only material you are permitted to process.
Ask for evidence and uncertainty. Verify against the original record, protocol or full text.
Make and record the human decision in your review system. Copy only a concise, verified summary to the worksheet.

Review the rules before screening

1 · Protocol and eligibility

Check protocol clarityUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Review this systematic review protocol section for ambiguity and internal consistency. Do not change the approved scope or invent decisions.

Review question: [paste]
Inclusion criteria: [paste]
Exclusion criteria: [paste]
Study designs and context: [paste]

Return a table with: wording to clarify, why it may lead to inconsistent decisions, a neutral question for the review team, and a suggested wording option. Mark suggestions as proposals requiring human approval.

Create a pilot calibration setUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Using only these eligibility criteria and the records I provide, identify which criteria may be difficult to apply consistently. Do not decide whether records are eligible.

Criteria: [paste]
Sample title/abstract records: [paste]

For each difficult criterion, explain what information is needed and propose a calibration question reviewers should discuss. Do not infer information absent from a record.

Use metadata to flag possible matches; verify in Zotero or your chosen system

2 · Deduplicate records

Flag possible duplicate pairsUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Compare these two bibliographic records and classify them as likely duplicate, likely distinct, or uncertain.

Record A: [paste complete metadata]
Record B: [paste complete metadata]

Compare title, DOI, authors, year, source and pagination. State which exact fields support the classification and what should be checked in the original source. Do not merge or delete records; a human must inspect the records.

Prepare a deduplication auditUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

From the supplied deduplication log, summarize the method and unresolved cases. Use only the log; do not add counts.

Deduplication log: [paste]

Return: matching rules used, manual verification performed, uncertain pairs retained or resolved, and fields still missing. Mark every count as “from supplied log” and flag arithmetic inconsistencies.

Use AI as a cautious second look, never as an independent reviewer vote

3 · Title and abstract screening

Provisional title/abstract decisionUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Apply the criteria below to each supplied title and abstract. For each record, give a provisional label: Include, Exclude, or Unclear. This is a preparation aid only, not a reviewer decision.

Inclusion criteria: [paste]
Exclusion criteria: [paste]
Records with IDs and abstracts: [paste]

For each record, cite the exact evidence from the supplied text, name the relevant criterion, and state missing information. Use Unclear when evidence is insufficient. Never invent facts or make a decision from title alone when the abstract is needed.

Compare conflicting reviewer rationalesUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Compare these two reviewer rationales against the agreed criteria. Do not cast a vote, choose a reviewer, or resolve the conflict.

Criteria: [paste]
Record ID and supplied abstract: [paste]
Reviewer A rationale: [paste]
Reviewer B rationale: [paste]

Identify the precise criterion or evidence in dispute and suggest a neutral question for human discussion. If the record does not resolve the issue, say so.

Ground every suggestion in the paper and record a primary exclusion reason yourself

4 · Full-text screening

Check a full-text record against criteriaUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Using only the supplied full text excerpts and eligibility criteria, prepare a provisional evidence table. Do not make the final inclusion decision.

Criteria: [paste]
Record ID and citation: [paste]
Relevant full-text excerpts with page numbers: [paste]

For every criterion, report: criterion, supporting/contradicting quotation or close paraphrase, page number, and status (met, not met, unclear, not reported). If evidence is absent, write “not reported”. Do not invent page references. End with questions for the human reviewer.

Check exclusion-reason wordingUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Review these proposed full-text exclusion reasons for clarity and whether each maps to one stated exclusion criterion.

Exclusion criteria: [paste]
Record IDs, proposed primary reason and evidence: [paste]

Flag reasons that are vague, overlapping, unsupported, or not a criterion. Suggest a clearer label, but do not change the record decision or create new criteria.

AI can check arithmetic; source exports and reviewer logs determine the counts

5 · PRISMA count reconciliation

Audit flow-count arithmeticUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Check the arithmetic and relationships in these review flow counts. Do not invent or fill in any missing count.

Counts copied from my verified log: [paste counts and definitions]

Show each calculation, identify inconsistencies or undefined categories, and list the source record I should inspect to resolve each issue. Keep records and reports distinct where relevant. Return “cannot verify” when the supplied information is insufficient.

Use the selected checklist; retain evidence and reviewer judgement

6 · Quality appraisal

Map appraisal evidence to checklist itemsUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Organize evidence from the supplied paper against the appraisal checklist. Do not assign an overall quality score or replace the reviewer’s judgement.

Checklist and response options: [paste]
Study ID and citation: [paste]
Relevant excerpts with page numbers: [paste]

For each checklist item, provide the exact supporting evidence and page, a tentative response only if directly supported, and what remains unclear. Use “not reported” when appropriate. Do not infer quality from the journal, author, or outcome.

Extract traceable facts first; synthesize only across verified included studies

7 · Data extraction and synthesis

Extract structured evidenceUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Extract data from the supplied included study into the fields below. Use only the supplied paper text. Do not infer or supplement missing facts.

Study ID: [ID]
Extraction fields: [paste field names]
Research question mapping: [paste RQs]
Paper text or relevant excerpts: [paste with page numbers]

Return one row per field with: extracted value, exact evidence or concise quotation, page number, mapped RQ, and confidence/uncertainty note. Use “not reported” when absent. Preserve distinctions in the paper.

Build an evidence synthesis matrixUse this as a cautious preparation aid; verify all suggestions against your source records.View prompt +

Copy and replace the bracketed fields

Organize these verified extraction rows to help compare studies. Do not introduce external facts or claim a pattern that is not supported by the supplied rows.

Research questions: [paste]
Verified extraction rows with study IDs and page references: [paste]

Create a matrix grouped by RQ and relevant context/method/outcome. Separate recurring findings, contrasting findings, and evidence gaps. For each synthesis statement, list the supporting study IDs and extracted evidence. Mark tentative patterns as tentative.

Use with any AI-assisted Stage 2 task

AI Output Verification Checklist

AI output is a working aid. Verify every decision and factual statement against protocol, reviewer logs, full text and source records. Tick items as you work; this checklist is not saved.

Criteria followedSuggestions map to the approved protocol and any change has human approval.Evidence traceableEvery paper-based statement has a verified record and page reference.Counts verifiedNumbers match original exports, decision logs and the defined PRISMA categories.Human decisions retainedReviewers made and documented screening, appraisal and extraction decisions.Missing data markedAbsent information is labeled “not reported” or “unclear”; it is not guessed.Privacy consideredOnly content permitted by institutional policy was entered into the AI service.

Do not treat an AI response as a reviewer vote, a source record, or a substitute for the audit trail.

Responsible-use instruction

Add this to a prompt when you need stronger safeguards:

Use only the information I provide. Separate direct evidence from suggestions. Do not invent citations, records, counts, quotations or page numbers. If information is missing, state “not reported” or ask me. Do not make final review decisions; a human reviewer will verify and record them.

Use only the information I provide. Separate direct evidence from suggestions. Do not invent citations, records, counts, quotations or page numbers. If information is missing, state “not reported” or ask me. Do not make final review decisions; a human reviewer will verify and record them.

Return to Stage 2 workspace →Browse tool guides